Scholarlink

<p>Mutations in Isocitrate dehydrogenase 1 (IDH1) create neoenzymatic activity that drives the oncometabolite 2-hydroxyglutarate, motivating selective small-molecule inhibitors. Here, we present a dual-strategy pipeline that integrates Bayesian neural network (BNN)-aided structure-based virtual screening (SBVS) with an active-learning-guided generative design loop. Beginning from ~ 3.1 million candidate structures, a BNN provides calibrated activity means and uncertainties that drive upper-confidence-bound acquisition, while a Transformer-based generative model proposes scaffold-diverse analogs optimized for predicted binding affinity, physicochemical constraints, and ADMET priors. Shortlisted compounds undergo consensus docking and triplicate 200-ns molecular dynamics (MD) per complex, followed by free energy decomposition and in silico ADMET profiling. We identify five chemically diverse leads (XS-1–XS-5) with stable binding modes and favorable predicted developability relative to AG-120. Residue-level analyses reveal context-dependent contributions—most notably His132, which exhibits high conditional ΔΔG despite lower contact frequency—supporting targeted SAR hypotheses. Our results demonstrate that coupling uncertainty-aware prioritization with goal-directed generation accelerates the discovery of selective mutant-IDH1 inhibitors while preserving chemical diversity and downstream viability.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Bayesian active learning-aided structure-based virtual screening reveals novel inhibitors of mutant IDH1

  • 期刊论文

  • 发表日期：

    2025年10月15日

查看全文
  • Sen Xu,
  • Yue Yang,
  • Chao Chen,
  • Xiaolong Lv,
  • Xiaojun Lei,
  • Haigang Wu,
  • Yuguang Lei

摘要

Mutations in Isocitrate dehydrogenase 1 (IDH1) create neoenzymatic activity that drives the oncometabolite 2-hydroxyglutarate, motivating selective small-molecule inhibitors. Here, we present a dual-strategy pipeline that integrates Bayesian neural network (BNN)-aided structure-based virtual screening (SBVS) with an active-learning-guided generative design loop. Beginning from ~ 3.1 million candidate structures, a BNN provides calibrated activity means and uncertainties that drive upper-confidence-bound acquisition, while a Transformer-based generative model proposes scaffold-diverse analogs optimized for predicted binding affinity, physicochemical constraints, and ADMET priors. Shortlisted compounds undergo consensus docking and triplicate 200-ns molecular dynamics (MD) per complex, followed by free energy decomposition and in silico ADMET profiling. We identify five chemically diverse leads (XS-1–XS-5) with stable binding modes and favorable predicted developability relative to AG-120. Residue-level analyses reveal context-dependent contributions—most notably His132, which exhibits high conditional ΔΔG despite lower contact frequency—supporting targeted SAR hypotheses. Our results demonstrate that coupling uncertainty-aware prioritization with goal-directed generation accelerates the discovery of selective mutant-IDH1 inhibitors while preserving chemical diversity and downstream viability.